Nonlinear Transform Source-Channel Coding for Semantic Communications

In this paper, we propose a class of high-efficiency deep joint source-channel coding methods that can closely adapt to the source distribution under nonlinear transform, it be collected name transform (NTSCC). considered model, transmitter first learns analysis map data into latent space, then transmits representation receiver via coding. Our model incorporates as strong prior effectively extract semantic features and provide side information for Unlike existing conventional methods, proposed NTSCC essentially both an entropy on representation. Accordingly, novel adaptive rate transmission hyperprior-aided codec refinement mechanisms are developed upgrade The whole system design is formulated optimization problem whose goal minimize end-to-end rate-distortion performance established perceptual quality metrics. Across test image sources with various resolutions, find method generally outperforms analog using standard classical separation-based digital transmission. Notably, potentially support future communications due its content-aware ability goal.